ARTEX: Exploited Open-Source Penetration Platform

Gerard Chukwu, CTI AI Technologist
TLP:CLEAR // CS-SA-26-1103

Executive Summary

The CRITICALSTART® Cyber Research Unit (CRU) is tracking the malicious use of ARTEX, an open-source, multi-agent penetration testing framework that uses large language models (LLMs) to automate intrusions from reconnaissance through data exfiltration.[1] Since at least August 2026, state-aligned espionage groups and financially motivated criminals have used ARTEX against organizations in financial services, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecommunications, and legal sectors.[1, 5] The framework uses no new exploits; it chains familiar weaknesses such as exposed credentials, over-permissioned cloud applications, and common web flaws at a pace that reached privileged cloud access in under three hours in some intrusions.[1, 2] On November 13, 2026, the Federal Bureau of Investigation (FBI) issued FLASH-20261113-001 warning that the developer's October removal of the source code has not stopped its spread.[1] Organizations should prioritize third-party application identity review, secret rotation, and behavior-based detection over indicator blocking.

Overview

ARTEX is a Go-based framework with an embedded Next.js web console.[5] Planning agents convert an operator's natural-language objective into tasks, worker agents run standard offensive tooling, and results feed back until the objective is met.[8, 9] The framework has no built-in model and connects to commercial or open-weight LLMs. Analysis of attacker-controlled instances shows DeepSeek Flash family models dominate because of low cost and few effective guardrails, with GLM, Grok, and Kimi also observed.[1, 6] The developer closed the source on October 8, 2026, citing misuse.[1] Forks, localized variants, archived copies, and a paid ARTEX PRO edition remain in circulation, and researchers have found backdoored ARTEX Docker images on public registries.[1, 9, 10]

Scope and Scale

The ARTEX threat landscape spans at least five publicly reported operator clusters, including China-linked espionage groups and Asia-based cybercriminal operators.[10] Its exposed infrastructure is substantial, with 1,465 internet-facing ARTEX consoles identified as of November 4, 2026. A separate assessment identified 1,608 active panels among 2,557 historically observed hosts.[6, 8] The largest reported espionage campaign targeted more than 200 organizations across at least 14 countries within 11 days, with post-compromise activity confirmed at nine organizations.[3] Separately, the largest supply chain campaign generated sign-in attempts against more than 2,250 Microsoft 365 tenants, peaking at 2,900 attempts per minute.[2] Attackers demonstrated the ability to obtain privileged cloud access in under three hours, while one suspected incident involved access to Azure Key Vault data just five minutes after a secret was exposed.[1, 10] These operations also required relatively modest resources, with reported LLM usage costs ranging from $145 to $1,500 per campaign.[6]

Attribution

No public evidence links these clusters beyond shared use of the tool by the following:[11]

  • VAPOR PANDA (CrowdStrike): China-nexus espionage group targeting foreign ministries, telecommunications, defense, semiconductor, and pharmaceutical organizations; separate from Storm-0940.[3]
  • Storm-0940 (Microsoft; overlaps GTIG's UNC7228): China-state aligned actor behind the ConnectWise Cloud Backup supply chain compromise.[2, 5]
  • Storm-3358 (Microsoft): Cambodia-based, financially motivated cluster targeting Southeast Asian financial organizations.[7]
  • Unattributed South Korea actor: financially motivated Chinese speaker (moderate confidence) whose campaign exposed personal data of about 67,000 people; related activity tracked as TAG-217.[1, 4, 11]

Timeline of Key Events

Date (2026)Event
JulyARTEX released as open source on GitHub[1]
Late AugustEarliest in-the-wild use identified in incident response engagements[5]
SeptemberStorm-0940 begins using ARTEX to automate attacks[2]
October 8Developer halts updates and closes the source[1]
October 17VAPOR PANDA campaign begins; ConnectWise disables seven Cloud Backup application identities[3, 11]
October 21Microsoft attributes the ConnectWise Cloud Backup supply chain compromise to Storm-0940[2]
October 29 to November 6CrowdStrike, GTIG/Mandiant, Censys, Check Point, and Microsoft publish campaign analyses[3, 5, 6, 7, 8]
November 13FBI issues FLASH-20261113-001[1]

Attack Methodology

ARTEX campaigns follow a conventional intrusion lifecycle. What changes is speed, parallelism, and persistence: agents retry failed steps within seconds and pursue multiple paths at once. Google Threat Intelligence Group (GTIG) recorded 27 cloud permission checks in 23 seconds during one intrusion.[5, 10]

Initial Access

Agents enter through whatever an internet-facing system exposes. The FBI lists unsigned Security Assertion Markup Language (SAML) assertions, XML External Entity (XXE) injection, Server-Side Request Forgery (SSRF), SQL injection, command injection, Server-Side Template Injection (SSTI), Insecure Direct Object Reference (IDOR), missing application programming interface (API) authentication, anonymous file upload, and hardcoded administrator passwords in JavaScript bundles.[1] ARTEX performs best against peripheral systems that receive less security attention, including partner portals, mobile backends, sales-support applications, and legacy services.[1, 12]

Identity and Cloud Abuse

Storm-0940 obtained credentials for ConnectWise Cloud Backup application identities, which already held consented permissions in customer Microsoft 365 tenants. The applications authenticated as themselves, so no user sign-in or Multi-Factor Authentication (MFA) prompt was involved.[2, 11] Where sign-in succeeded, agents ran Graph reconnaissance, added application credentials, and assigned privileged roles.[2] Across Azure and AWS, agents also exfiltrated Key Vault data, ran scripts through VM Run Command, and searched DevOps and CloudWatch logs for secrets.[10]

On-Premises Post-Exploitation

GTIG and Mandiant observed agents enumerate more than 110,000 Active Directory objects, Kerberoast service accounts, read Local Administrator Password Solution (LAPS) passwords, and use DCSync to obtain the krbtgt key.[1, 5] Agents deployed JSP web shells, sent output to 491 rotating webhook.site URLs, and fetched the WISP, vshell, and AdaptixC2 remote access tools.[7, 10, 11] VAPOR PANDA directed its agents to find Windows hosts with no Endpoint Detection and Response (EDR) sensor, then deployed the BLOODALCHEMY backdoor through Dynamic Link Library (DLL) sideloading by a renamed GameHook binary running as svchost.exe.[3, 10]

MITRE ATT&CK Mapping

The mapping below consolidates techniques reported by the FBI, CrowdStrike, Microsoft, and GTIG/Mandiant.[1, 3, 4, 5]

IDNameObserved Activity
T1588.007Obtain Capabilities: Artificial IntelligenceAcquisition and deployment of ARTEX and commercial or open-weight LLMs[4]
T1595Active ScanningDirectory fuzzing of 500+ requests and high-frequency endpoint enumeration[1]
T1580Cloud Infrastructure DiscoverySub-second permission checks against Azure and AWS credentials[1]
T1190Exploit Public-Facing ApplicationSQL injection, XXE, SSRF, SSTI, command injection, and IDOR against web applications[1]
T1195Supply Chain CompromiseCompromise of a multi-tenant vendor application[1, 2]
T1199Trusted RelationshipUse of pre-consented vendor application access to downstream tenants[1, 2]
T1078.004Valid Accounts: Cloud AccountsAuthentication as compromised application identities[1, 11]
T1110.003Brute Force: Password SprayingCross-tenant spraying peaking at 2,900 failures per minute[2]
T1552.001Unsecured Credentials: Credentials in FilesHarvesting secrets from public repositories and web roots[1, 10]
T1552.005Unsecured Credentials: Cloud Instance Metadata APISAML and XXE chain used to query the metadata service[1, 11]
T1606.002Forge Web Credentials: SAML TokensAbuse of unsigned SAML assertions; forged admin tokens using unrotated keys[1, 11]
T1505.003Server Software Component: Web ShellNumbered JSP and PHP web shells, including via MySQL log poisoning[1, 11]
T1098.001Account Manipulation: Additional Cloud CredentialsCredentials added to existing OAuth applications[1, 2]
T1098.003Account Manipulation: Additional Cloud RolesPrivileged role assignments after authentication[1, 2]
T1558.003Steal or Forge Kerberos Tickets: KerberoastingService ticket requests against Active Directory[1, 5]
T1003.006OS Credential Dumping: DCSyncReplication of the krbtgt key and domain controller hashes[5, 11]
T1518.001Software Discovery: Security Software DiscoveryInventory of EDR coverage to find unprotected Windows hosts[1, 3]
T1574.001Hijack Execution Flow: DLLBLOODALCHEMY sideloaded by a renamed GameHook binary[1, 3]
T1036.005Masquerading: Match Legitimate Name or LocationGameHook binary renamed svchost.exe[1, 10]
T1105Ingress Tool TransferRetrieval of WISP, vshell, and AdaptixC2 with curl and netcat[1, 7]
T1555.006Credentials from Password Stores: Cloud Secrets Management StoresAzure Key Vault exfiltration[1, 10]
T1530Data from Cloud StorageMass blob reads and SharePoint and OneDrive downloads[1, 10]
T1114.002Email Collection: Remote Email CollectionMailbox access through Microsoft Graph[1, 10]
T1567.004Exfiltration Over Web Service: Exfiltration Over WebhookCommand output sent to rotating webhook.site URLs[11]
T1090.003Proxy: Multi-hop ProxyResidential proxies, Tor, and SOCKS relays to rotate source addresses[1, 4]
T1572Protocol TunnelingTunneling through Chisel and SoftEther[1]

Relevant Vulnerability Context

ARTEX has no Common Vulnerabilities and Exposures (CVE) identifier or patch, and the FBI cites vulnerability classes rather than specific CVEs.[1, 11] Operator infrastructure did hold exploit builds for two Linux flaws, shown with Common Vulnerability Scoring System (CVSS) scores.[10]

Implications for Organizations

Organizations must account for attack timelines measured in minutes, not days. Microsoft documented ARTEX-enabled intrusions that progressed from initial targeting to privileged cloud access in under three hours, while another investigation observed an AI agent move from exposed GitHub secrets to Azure Key Vault data in just five minutes.[1, 10] Automated agents can scan, exploit, validate credentials, and pursue multiple attack paths with limited human intervention. Alert generation alone is insufficient if security teams cannot rapidly investigate and revoke compromised access. Third-party application integrations and OAuth permissions also represent significant exposure. The ConnectWise compromise demonstrated how stolen application credentials could provide a path into more than 2,250 downstream Microsoft 365 tenants without requiring interactive user authentication.[2] Organizations should therefore treat application identities, service principals, OAuth grants, and third-party integrations as privileged access, with particular attention to excessive permissions and credentials that remain valid after a vendor incident.

ARTEX-related activity does not, by itself, establish malicious intent or identify the operator. Researchers have observed the same open-source frameworks across espionage, financially motivated operations, and legitimate penetration testing engagements.[6, 10] The growing availability of exposed consoles, modified forks, and alternative frameworks such as PentAGI and CAIRN further complicates attribution and detection. Organizations should focus on observable behaviors, including automated reconnaissance, password spraying, unauthorized OAuth application changes, credential harvesting, privilege escalation, and unusual Microsoft Graph activity, rather than relying solely on ARTEX-specific signatures. Given the low reported operating costs and ability to reuse or modify open-source tooling, CRU assesses with high confidence that adoption of agentic offensive frameworks will continue expanding through 2027.

Prioritized Mitigation Strategies

To reduce the risk of ARTEX-enabled intrusions and limit the impact of automated attacks against cloud and enterprise environments, CRU recommends the following actions:

Immediate (Within 24 Hours)

  • Identify compromised third-party applications: Search every Microsoft 365 tenant for the ConnectWise application IDs listed in Appendix A of the full advisory (PDF). Search by application ID rather than display name, keep affected connections disabled pending vendor guidance, and review associated service principal sign-in activity from October 1, 2026, onward.[1, 11]
  • Audit cloud identity changes: Review Microsoft Entra ID audit logs dating back to August 1, 2026, for unauthorized application credentials, ownership changes, consent grants, and privileged role assignments. Investigate authentication errors AADSTS7000215 and AADSTS7000222, particularly when repeated failures are followed by a successful sign-in.[1, 10]
  • Hunt for ARTEX artifacts: Search endpoint, network, and cloud telemetry for the indicators in Appendix B of the full advisory (PDF), including ARTEX-related user-agent strings, paths beginning with artx, and files matching artex-probe-*.[9, 10, 11]
  • Preserve forensic evidence: Retain Microsoft Entra ID, Microsoft Purview Audit, and SIEM logs before disabling credentials or removing suspicious access. Avoid immediately deleting affected enterprise applications, as doing so may destroy evidence needed to investigate the scope of compromise.[1]

Short-Term (Within Two Weeks)

  • Eliminate exposed credentials: Scan source-code repositories, CI/CD pipelines, infrastructure-as-code files, and container images for embedded credentials and LLM API keys. Rotate exposed secrets, invalidate associated tokens, and enable GitHub secret scanning and push protection.[1, 10]
  • Reduce internet-facing attack surfaces: Inventory externally accessible applications, including partner portals, mobile backends, and legacy services. Restrict administrative interfaces through VPN or zero-trust access controls, disable anonymous file uploads, remediate SSRF vulnerabilities, enforce IMDSv2 on AWS workloads, and reject unsigned SAML assertions.[1, 10, 11]
  • Strengthen endpoint and identity protection: Verify EDR coverage across Windows and Linux servers, require phishing-resistant MFA for user accounts, and apply appropriate Conditional Access and workload identity controls to limit unauthorized authentication and privilege escalation.[3, 9, 11]
  • Monitor suspicious outbound activity: Investigate unexpected connections from enterprise servers to LLM API providers, including api.deepseek[.]com and open.bigmodel[.]cn, as well as webhook.site and other request-capture services. Correlate these connections with unusual process execution and outbound data transfers.[1, 11]

Structural

  • Automate identity containment: Develop response playbooks that can disable compromised service principals, revoke application credentials, and invalidate associated access tokens. Require analyst review where necessary to prevent disruption to legitimate business services.[9]
  • Govern third-party application access: Treat multi-tenant application permissions and OAuth consent grants as privileged access. Enforce least privilege, periodically recertify application permissions, remove unused integrations, and establish appropriate rate limits for business-critical APIs.[1, 10]
  • Establish visibility into authorized AI security testing: Require internal red teams and approved security providers to register their use of ARTEX and other agentic penetration-testing frameworks with the SOC. Document approved infrastructure, testing windows, and expected activity so analysts can distinguish authorized assessments from malicious operations.[9]
  • Improve detection of automated intrusions: Develop behavioral detections for high-frequency reconnaissance, repeated authentication attempts, rapid permission enumeration, unexpected service principal modifications, and unusual access to cloud secrets and storage. Correlate activity across identity, endpoint, and cloud environments rather than relying exclusively on ARTEX-specific indicators.

Conclusion

ARTEX demonstrates how agentic AI is changing the speed and scale of cyber intrusions, allowing threat actors to automate reconnaissance, exploit existing weaknesses, and expand access across cloud and enterprise environments with limited human intervention. The risk extends beyond organizations directly targeted by attackers: compromised third-party application identities and excessive OAuth permissions can expose entire downstream customer environments. As ARTEX forks and similar frameworks become more widely available, organizations should expect continued adoption by both state-aligned and financially motivated actors. CRU recommends prioritizing visibility into application identities and third-party integrations, rapid credential revocation, and behavioral detection capable of identifying automated attack activity across cloud and on-premises environments. The ability to detect and contain compromised identities in minutes, rather than relying on traditional investigation timelines, will be critical to limiting the impact of agentic attacks.

For more threat reports, including OPTIMJACK and Java4Shell, visit Critical Start's Intel Hub. This advisory was written using the best intelligence available at the time and is subject to change as additional information becomes available.

What Critical Start is Doing

The CRITICALSTART® Cyber Research Unit (CRU) disseminated a brief of the threat as part of the 10-22-2026 and 11-05-2026 CTI Updates and is providing this Security Advisory as a follow up. CRU will continue to monitor ARTEX developments, as well as Agentic AI-accelerated attacks while working closely with the SOC and Security Engineering team to implement any relevant detections and increase vigilance. Critical Start customers are encouraged to reach out to support@criticalstart.com for any further enquiries.

To learn how Critical Start can help you detect, investigate, and respond to threats faster, connect with a Critical Start team member.

Get Started →